As desirable as it might be to predict early in the course of coma whether a patient will do well or poorly, all studies of coma prognosis are plagued by inherent methodologic problems that tend to diminish the utility of the derived criteria: especially the tendency of poor prognoses to be self-fulfilling, the rapid drop-off in patient population due to death from nonneurologic causes, and the need to lump, for the sake of statistical significance, outcome categories that ought to be kept distinct for purposes of ethical decision making. Even for a methodologically ideal study, if 100 per cent of the N patients fulfilling a particular criterion experienced the same poor outcome, the probability of a false prediction of poor outcome in the next patient meeting that criterion is approximately 1/(N + 2), which is hardly negligible for a realistically sized study. Moreover, there is a 50 per cent chance of at least one false-positive prediction among the next (N + 1) patients fulfilling the criterion. Given this inherent unreliability of early predictors for individual patients, given that decisions to continue life support are reversible, whereas decisions to withdraw it are usually not, and given that the death of a patient with potential for recovery is a more serious error than the (typically) transient prolongation of life of a patient destined soon to die anyway or (much less commonly) to remain in a chronic vegetative state, it would seem prudent to continue life support for all patients during the first few weeks or months of coma or vegetative state, regardless of early indicators of poor outcome. Early prognostication can still serve other useful purposes, however, including the counseling of families, triage and DNR decisions, and future clinical investigations of brain-resuscitative measures.